Beyond SHAP: Reliable feature selection methods for clinical prediction models.

Area of Science:

• Clinical prediction modeling
• Machine learning interpretability
• Biostatistics

Background:

• Model-dependent feature importance methods in clinical prediction lack validation, leading to unreliable insights.
• Algorithm selection significantly impacts feature rankings, even with comparable prediction accuracy.
• Existing validation focuses on accuracy, neglecting the crucial aspect of feature importance reliability.

Discussion:

• Proposes a novel framework combining statistical and information-theoretic approaches for robust feature importance assessment.
• Integrates monotonic relationship detection (Spearman's correlation, Kendall's tau) with p-value assessment.
• Incorporates complex interaction analysis using Mutual Information and Effective Transfer Entropy.

Key Insights:

• Algorithm choice critically affects feature importance, irrespective of predictive performance.
• Current validation practices for clinical prediction models are insufficient.
• The proposed dual methodology offers a more reliable assessment of feature importance.

Outlook:

• Enhances the trustworthiness and interpretability of clinical prediction models.
• Facilitates more accurate identification of key clinical predictors.
• Aims to improve clinical decision-making through reliable model insights.